I was obviously intrigued and investigated the ChatGPT software. The latest 3.5 version of the software was launched by OpenAI in November 2022. OpenAI is a for-profit software firm that has been researching the field of artificial intelligence (AI) with the stated goal of developing friendly AI. It’s interesting that friendly is a key part of their mission statement because many AI industry pundits predict that AI will likely eventually compete with humans for resources, much like Skynet in the Terminator movies.
ChatGPT is written atop OpenAI’s third generation of software and is aimed at communicating in a written or conversational way so that a reader can’t tell the difference between the software and a human. The company has numerous investors, but Microsoft just offered to buy a 49% stake in the company for $10 billion. This instantly has me wondering when there will be a fee to use the software instead of the free version that is available now.
The press on ChatGPT has been over-the-top. I’ve seen articles comparing the impact of the launch of ChatGPT to other big events in web history, like the first web browser or the iPhone. Articles are touting that the software will mean that programmers will no longer have to write code, that students will no longer have to write papers, and that there will soon be no need for journalists (or bloggers!)
Early-generation AI writing software has been around for a few years and many baseball box scores and press summaries of quarterly earnings reports have been generated by software. These are writing tasks that are formulaic and repetitive, and I doubt that most folks noticed – although the software never captured the magic of a sports reporter like Shirley Povich, who I enjoyed reading every day for years in the Washington Post.
I had to give this platform a try. Was this software capable of writing something like this blog? If so, it would make me reconsider writing every day because if the software is that good there won’t be much need for human writers before long. As I was testing, I also considered the idea of using the software to get a jump start on a new piece of writing – the idea of seeing if the software could structure and organize an idea would be a time saver if the results were usable.
You can give complicated instructions to the software. You can provide the topic, the desired length of the end product, and describe the desired style of writing. I gave the software several topics to write about, and I was impressed with the speed of the process. The finished product is created almost as soon as you say go to the software.
But I was underwhelmed by the results. The sentences are grammatically perfect, and each paragraph has a topic and tries to make a point. Yet the end result was stilted, and some paragraphs were unreadable – I had to reread them several times to try to decipher the point (but for all I know, my readers have to do the same thing!).
The biggest flaw was that the writing was full of factual errors. That makes a lot of sense because the software distills what is written on the web when writing content. It takes the good and the bad, the factual and non-factual, and the easy-to-understand and obtuse writing that exists on the web and mashes it in a synthesis of what it finds. I realized that I would have to fact-check everything ChatGPT writes because the software has no way to discern what is true or untrue. There is a term for this among data scientists, and I read that ChatGPT currently has a hallucination rate of between 15% and 21%, meaning that it seems to make up that percentage of facts in its writing.
I know there is instant hope among students that this software can churn out the dreaded school essay – but that doesn’t look likely. The software has been out for only two months, and I saw that a software engineer has already developed a program that can detect with more than 90% accuracy if something is written by a human or by ChatGPT. Students beware.
The day will likely come when the ChatGPT writing gets better, but there is nothing in this software today that would make me consider giving up writing or even using this as a tool. The hallucination rate means I can’t trust it to be factual, so it’s not even worth using to create a kernel of an idea for a blog. Most importantly, the output is not readable – it’s all perfect English, but I couldn’t understand the point of about half of what it wrote for me. If my blogs are going to be unreadable, I want the obtuseness to be fully human-generated!